Visual Inspection Of Bare Driver Circuit Board Based On Convolution Network

With the country's incentive support for high-end manufacturing industry,enterprises have more and more demand for servo system products,and the strict control in the production process is conducive to the improvement of product quality.As an important part of servo system,the driver circuit board visual detection is one of the key parts to improve the product quality.In order to improve and enhance the efficiency,accuracy and robustness of the visual inspection process of the driver circuit board,this paper takes the improvement of the detection technology in the production process of the driver product as the research background,takes the bare driver circuit board as the research object,and makes an in-depth study on the matching algorithm and recognition algorithm in the visual detection process of the bare circuit board,and designs and implements a bare driver circuit board vision The main research contents of the system are as follows:Aiming at the problem of fast and accurate matching and registration of bare circuit board based on calibration points,a feature description method and similarity measurement method based on azimuth environment features are proposed for bare board matching and registration.Firstly,this method uses image processing to find the preset calibration points on the bare board.Then,according to the azimuth environment feature description method and feature similarity measurement method based on Chi square distribution,their feature vectors are calculated and matched,so as to realize the matching of calibration points of template image and test image and the registration of final bare board image.Aiming at the problem of fast and accurate matching and registration of bare circuit board based on local feature points,a learning description network matching method of global negative sample mining is proposed for bare board registration.This method uses the global negative sample mining strategy of this paper to train a local description network of L2 Net with the existing data set,to describe the feature points detected by Fast method,and to match these feature points with the similarity measurement method of Euclidean distance,so as to calculate the registration of affine transformation matrix and final bare board.In order to compare the performance improvement of this training strategy to the network,we use a variety of evaluation methods to compare with other description methods,which shows that the proposed training strategy has a certain performance improvement to the network,which can more accurately match the local feature points,so as to improve the robustness and accuracy of the bare board registration.Aiming at the current bare circuit board defect identification using traditional empirical rule judgment,this paper proposes a method of bare circuit board defect recognition based on reduced VGG convolution network,which is used to detect and recognize the six major defects of bare driver circuit board,such as missing hole,mouse bite,open circuit,short circuit,burr and copper leakage.Firstly,this method compares the difference between the template image and the test image after the local adaptive threshold,then uses morphology and contour tracking to extract the defect image.Then,this method trains a reduced VGG network to identify the defect of the driver bare board,and the experiment shows that this method is more accurate than the traditional method in identifying the defect of the driver bare board.Finally,a hardware and software platform of bare driver circuit board visual inspection is designed and implemented.The function,selection,design and development of each hardware and software module of the system are described in detail.The learning description network matching method and the defect identification convolution network proposed in this paper are implemented in the system to realize the on-line accurate detection of bare driver circuit board.